CUA-Master

CUA-Master is a research codebase that builds training data and trajectories for computer-use agents (AI that operates screens) for researchers and developers.

What it is good for

  • Turning instructional videos into training data for computer-use agents
  • Generating verifiable tasks to test agents that operate a computer
  • Collecting agent trajectories in sandboxed virtual machines
  • Converting agent trajectories into fine-tuning datasets

At a glance

Who it is for
Setup difficulty
Developer ยท Requires programming

Level 5 of 5. You need to write code to use it.

Cost
Free and open source
Type
Framework or SDK
Tasks
Works with
ms-swift, Hugging Face

How to get started with CUA-Master

  1. Install the package from the repository with the pip install -e ".[seeding,annotator,collection,dev]" command.
  2. Copy .env.example to .env and fill in your API keys and sandbox service endpoints.
  3. Run the mock collection smoke test with cuamaster-collect using configs/collection/mock_command.example.json, which needs no VM.
  4. Use --help on any command to see its options.

From the project's own documentation. If a step does not match what you see, the official pages have the latest instructions.

Before you try it

  • Requires API keys and sandbox service endpoints set in a `.env` file.
  • You must plug in your own VM or sandbox backend for real trajectory collection.
  • Aimed at researchers building training data, with a code-based workflow rather than a ready-to-use assistant.

Momentum

Rank this week
Not ranked (last #19, week 45, 2026)
Best rank
#19 week 45, 2026
GitHub stars
128

GitHub numbers as of Nov 16, 2026. Updated weekly.

22 79 136 Nov 1, 2026 Nov 16, 2026
GitHub stars for CUA-Master: 30 on Nov 1, 2026, 128 on Nov 16, 2026. Source: GitHub.

Ranking history

First ranked in week 45, 2026.

Questions

Is CUA-Master free?
Yes. It is free and open source.
Do I need to know how to code to use CUA-Master?
Yes. You need to write code to use it.
What does CUA-Master work with?
According to its own documentation, it works with ms-swift and Hugging Face.
How do I get started with CUA-Master?
1. Install the package from the repository with the pip install -e ".[seeding,annotator,collection,dev]" command. 2. Copy .env.example to .env and fill in your API keys and sandbox service endpoints. 3. Run the mock collection smoke test with cuamaster-collect using configs/collection/mock_command.example.json, which needs no VM. 4. Use --help on any command to see its options.
What should I know before trying CUA-Master?
Requires API keys and sandbox service endpoints set in a `.env` file. You must plug in your own VM or sandbox backend for real trajectory collection. Aimed at researchers building training data, with a code-based workflow rather than a ready-to-use assistant.
What is CUA-Master good for?
Turning instructional videos into training data for computer-use agents, generating verifiable tasks to test agents that operate a computer, collecting agent trajectories in sandboxed virtual machines, and converting agent trajectories into fine-tuning datasets.

Similar agents to CUA-Master

  • Agent-S

    Is a ready framework that operates the computer directly, rather than a pipeline for generating training data.

  • OpenSandbox

    Provides the isolated sandbox environments that CUA-Master expects you to supply yourself.

  • E2B

    Offers hosted cloud sandboxes for agents to run code, which could serve as a backend for collection.

  • agent-desktop

    Command-line tool for agents to operate desktop apps, with no data synthesis or training pipeline.

  • rea

    rea is an MCP server (a connector that lets AI apps use another tool) that helps developers work out how an app or binary behaves without its source code.

  • Folkbench

    Folkbench is a desktop app and public ranking site that compares AI API relay services and switches coding tools between them for Claude Code and Codex users.

Last updated Nov 16, 2026. Tracked since Nov 1, 2026. Parts of this summary required judgment because the source material was brief. Spotted an error? Tell us.